Implementation And Experimentation of Coherent Inference in Game Tree

Abstract

The game tree problem has been a focal and intriguing area for decades; people are fascinated by the idea that a computer could have rational thought. In 1950, Shannon's paper[1] presents a simplified mathematical procedure enabling a modern computer to play chess, since then a numerous of game tree search algorithms have emerged. Even though great progress has been made, there is no general algorithm that can perform perfectly in complex strategy games like Go. In this thesis, a coherent inference game tree search algorithm[2] showing some potentials in artificial game trees is analyzed in detail.It presents an approximate mechanism to infer off-policy optimal values of nodes in the game tree given on-policy scores from random roll-outs. Theoretical foundation and techniques employed in this algorithm are elaborated. Furthermore, experiments are carried out to examine the performance of this algorithm in real game tree and show its pros and cons.